Introduction
In the past few months, one of our primary research goals has been to develop a frontier model, Dawn 1, to power our always-on mental health platforms. This is a massive ongoing undertaking, with dedicated initiatives on data, infrastructure, and evaluation, and significant input from clinicians along the way. Today, we're outlining our progress on post-training GLM-5.3 into Dawn 1, a model that is considerably more capable and cost-efficient, while being safer and less sycophantic.
General performance and cost-efficiency
Because mental health does not follow a schedule, nor does it look the same for any two people, we firmly believe in proactive and personalized care. As such, Dawn 1 isn't just a chatbot role-playing a therapist. When tied to our harness, it proactively reaches out in the right moments, continuously holding a memory and clinical profile of each one of our members. During inactive periods, Dawn 1 actively reflects on conversations, finding ways to personalize the journey going forward.
And while solid performance can be achieved with the right harness,1 we find strong synergies with post-training that allow for pushing the frontier with continuous learning from our unique data. We take the harness and the model as two components that should be developed together, and adapt our evaluation and training approaches accordingly.2 Many of our RL environments thus focus on aspects like instruction-following, memory usage, and situational awareness. By adapting the model's native reasoning efforts, Dawn 1 becomes more capable than its backbone at a fraction of the tokens.3
Figure 1. Mental health intelligence-cost frontier. Mental health intelligence index is a combination of benchmark scores for several capabilities like memory usage, situational awareness, safety, etc. Our post-trained model is both more intelligent and less verbose per turn.
Sycophancy
We believe optimizing our model to the aforementioned capabilities leads to more engagement, which is a predictor of positive outcomes in healthcare.4, 5 Unfortunately, AI-based mental health support is often engaging because it is sycophantic: models reinforce user's perspectives, apologize unnecessarily, over-praise, and offer too little pushback.6, 7 Such behaviours may reinforce delusional beliefs and contribute to what has been described as "AI psychosis."8–10 As such, one of our primary goals in training is to reduce the deeply ingrained tendency of frontier models for sycophancy.
Figure 2. Sycophancy rates relative to our model's backbone. Our post-training process significantly reduces sycophantic tendencies on contexts with real users.
Safety
Safety is the most important dimension in our training process. Mental health is a high-stakes application, and we know our model may face situations where members exhibit the potential for self-harm, harm to others, and other concerning behaviors. We don't depend exclusively on our model for safety; in fact, in our most recent NeurIPS publication, MindGuard, we detail the system that inspired one component of our safety layer. That said, in-house clinicians have created specific safety protocols for high-risk situations. These protocols outline how the model must behave based on specific user actions, meaning they must be enforced across conversations rather than isolated turns. This means we had to design and support multi-turn environments to optimize for safety during RL. Each rollout simulates different user behaviours and rewards are given at both turn and conversation levels.
Figure 3. Multi-turn safety protocol adherence rates. RL consistently improves our model's ability to follow our safety protocols, surpassing frontier proprietary systems.
In the rest of this post, we provide more detail into some of the data, evaluation, and infrastructure work we did to create Dawn 1.
Evals
What constitutes good mental health care is highly subjective, even among clinicians. In the context of Dawn, we split the problem in the following dimensions, ordered by verifiability:
- Harness adherence: Dawn 1 must behave according to our harness specifications (e.g., tools, proactive and reflexive flows, instructions).11
- Core capabilities: Dawn 1 must be safe, clinically adequate, and display strong memory usage and situational awareness (e.g., temporal awareness, common sense).
- Feel: Dawn 1 must feel like a unique, engaging experience, without being sycophantic.
We strive to design challenging, realistic evals in a swiss cheese fashion, focusing on single aspects from the categories above, and taking them together when assessing the performance of any model iteration. We also leverage pre-made evaluations on tasks like code and clinical adequacy (e.g., MindEval), as we find these capabilities correlate moderately with the model's capacity to adapt to on-the-fly, fast production harness optimizations.12
Our first priority when creating an eval is realism, so we begin by automatically finding evaluation opportunities in anonymized conversations taken directly from our app.13 For example, we can directly use contexts with protocols triggered by our safety classifiers for evaluating how safe our model is across several turns. Uncovering opportunities to evaluate aspects like memory usage or sycophancy, however, is not as trivial. In those cases, we have used LLMs to classify user messages according to whether they would elicit specific capabilities in the following response(s), and we are continuously improving our methods for extracting the most value from our data. These are the steps that ensure a virtuous flywheel between deploying and continuously improving our models.
However, production data is inevitably biased to the sub-population our users represent,14 so we also leverage synthetic data to broaden test coverage. Some scenarios are simply too rare to find in production, or we may want to test our model before it faces certain situations (e.g., interacting with high-risk patients).15
Below, we showcase examples from some of our memory, situational awareness, and safety evaluations, with real completions from GLM 5.3 and Dawn 1.
Figure 4. One scenario from each of our memory, situational-awareness and safety evaluations, answered by GLM-5.3 and Dawn 1, with the judge's verdicts. Replies are real completions, trimmed where marked (...).
Besides aspects directly related to Dawn, our research is also focused on finding and solving fundamental challenges with automatic evaluation, such as self-preference bias in rubric-based evaluation (see our recent COLM paper).
Training
For creating training data we follow the same principle as evaluation: where possible, we leverage anonymised real data instead of synthetic one. This raises the challenge of emulating the production environment during rollout generation to minimize training-production mismatch (much like in Cursor's Composer 2 technical report). In practice, we have to, at least, restore environment and tool states, tool definitions, and sometimes simulate member behaviour realistically. To this end, we have built and continue to develop Dawn Gym, a codebase for simulating conversations as faithfully as possible to production, which we use directly as training and evaluation harness. Dawn Gym is yet another piece in our paradigm of harness and model co-design.
Furthermore, to ensure harmony between harness and model development, we train Dawn 1 to be versatile. To this end, among other interventions, we train Dawn 1 on what we call "side environments". These typically involve tasks that are part of our harness but that Dawn 1 is not necessarily expected to perform. For example, we have a side agent that helps Dawn 1 decide practical interventions to suggest to a member at certain points. Dawn 1 normally observes the outputs of this agent as a tool response, but we have an environment for training Dawn 1 on the task itself.16
We embed clinical expertise on most of our in-domain environments, be it in the form of expert-designed rubrics, or even judges trained on actual clinical annotations and preferences.
Below, we showcase the training process in 3 of our environments: The Mill multi-turn, where we simulate multi-turn continuations to production conversations; intervention recommendation, the environment described in the previous paragraph; and an environment that rewards leveraging psychology knowledge during specific moments of real conversations.
Figure 5. RL training progress on three of our environments.
Infrastructure
Training Dawn 1 required plenty of infra-related efforts and, in the end, was only possible by relying on the efforts from the open-source community and its artifacts as a base for our post-training stack.
Starting from Prime-RL, we rework several of its components, making custom-fit changes for our NVIDIA B300 Blackwell cluster. These range from fixing out-of-the-box NCCL HCA rail affinity issues, evenly distributing orchestration and checkpointing across trainer processes, incorporating and revamping specific GLM kernels across the MLA and MoE modules and finally, reducing GPU memory pressure by relying on Muon and the latest Dion3 implementation. Combined, these changes enable the quick iteration speed required to produce a strong model in such an undocumented domain, as seen below.
Figure 6. Cumulative GLM training speedup and memory requirements on 64 B300 GPUs. Reference is an extrapolation from smaller model shape due to OOM; the dashed line marks GPU memory capacity.
Next steps
Dawn 1 is a first milestone, and most of the work described above is still very much in progress. Below, we outline two directions we are particularly focused on.
Recursive self-improvement
Recursive self-improvement has become something of a hyped topic, but parts of the concept have very practical applications for us. By having our model analyse anonymized production conversations ingested by The Mill, we have been able to exploit its clinical intelligence to surface training opportunities in both the harness and the model itself. Quirks like those we describe in Appendix A, for instance, can be spotted automatically and at scale. Tracing them back often shows that the line between model and harness is blurry: "early sign-offs", where the model prematurely wraps up a conversation, are both a quirk of the backbone and a behavior provoked by our harness.17 We're also betting on something more ambitious: because Dawn's clinical intelligence surpasses that of general-purpose models, it can help improve its own future iterations in ways other models cannot. This becomes especially relevant as we also train Dawn to detect clinical issues, since every gain in that ability directly strengthens its capacity to critique its own behaviour. Naturally, our clinicians remain in the loop, validating the issues the model surfaces before they inform training.18 We will detail our progress on this front in a future blogpost.
Evaluation and user simulation
Evaluation remains one of the most challenging areas of our work. Due to the open nature of mental health support, it is still hard to identify the full surface area we care about, and especially to connect it to healthy engagement rather than the sycophantic kind we discuss above. User simulation goes a long way for certain aspects, such as testing safety protocols, memory usage across several turns, or scenarios that are very relevant yet infrequent in real data. Other behavioural aspects, however, are much harder to simulate faithfully, such as a member gradually disengaging or dropping off altogether.19 As our RL environments increasingly require multi-turn rollouts, realistic user simulation becomes ever more central, both for evaluation and for training. Improving how we simulate members is therefore one of our main research streams.
Appendices
Appendix A: Mitigating pre-existing GLM 5.3 quirks with RL
We find that a side-effect of training Dawn 1 to be compliant with its harness is that we mitigate some pre-existing quirks in GLM 5.3. These quirks are rare, but can lead to errors in production and to a worse user experience due to increased latency and incoherent text. We highlight a few examples in the plots below, like instances where the model's generation contradicts a preceding tool result, or even its own reasoning. We have several mechanisms in place to mitigate these behaviors, and manage at times to decrease their prevalence almost tenfold.
Figure A1-5: How often GLM-5.3's quirks show up over the RL run behind Dawn 1, each relative to its rate when RL starts (= 100%).
